LukeLin
LukeLin
> Disabling all CUDA extensions seems to have done it. It seems not a good solution. I wonder whether now we have a good cuda formatter.
> hey @LukeLIN-web, re: > > > AttributeError: 'Constant' object has no attribute 'lineno'. > > It's already fixed in mainline by #3201. > > I don't know about the...
> > I still has this problem AttributeError: 'Constant' object has no attribute 'lineno'.我仍然有这个问题 AttributeError： 'Constant' object 没有属性 'lineno'。 > > With which version? It was not patched in Triton...
打算做做, 目前应该是两个问题：算子和kvcache的内存管理模块.
目前的想法: 算子可以调用flashinfer.
For Unet, you could refer https://github.com/MrYxJ/calculate-flops.pytorch/issues/23#issuecomment-2522025272
It is not easy, we need to change CNN batch to a dict. ``` if isinstance(input_sample, dict): _ = flops_model(**input_sample) else: _ = flops_model(input_sample) ``` and change a lot of...
I thought it is much more than 60 000 ... Dataset Size = 2135488 Examples ,why n_train_examples=len(train_dataset)
what is gxx usage here ?
Currently cannot do it, because we create setup.py from setup.py.cfg